Local Dependence in Latent Class Analysis of Rare and Sensitive Events

Area of Science:

• Survey Methodology
• Statistical Modeling
• Psychometrics

Background:

• Latent class analysis (LCA) is crucial for survey methodologists to assess measurement error, evaluate survey methods, and estimate population prevalence bias.
• LCA is applicable when gold-standard measurements are unavailable, provided indicators meet identifiability criteria.
• Model validity in LCA is threatened by local dependence, stemming from bivocality, behaviorally correlated error, or latent heterogeneity.

Purpose of the Study:

• To examine threats to local independence in LCA, including bivocality, behaviorally correlated error, and latent heterogeneity.
• To provide insights into questionnaire designs that minimize local dependence.
• To explore modeling strategies for mitigating the effects of local dependence on statistical inference, particularly for rare and sensitive outcomes.

Main Methods:

• Investigated three potential causes of local dependence: bivocality, behaviorally correlated error, and latent heterogeneity.
• Focused on developing and testing a practical approach for diagnosing and mitigating model failures in LCA.
• Empirically tested the proposed modeling strategy using real-world data from a national survey on inmate sexual abuse, where measurement errors are a significant concern.

Main Results:

• The proposed modeling strategy demonstrated success in reducing local dependence bias in estimates.
• The effectiveness of the strategy varied based on the quality of indicators used in the analysis.
• Even with the proposed strategy, local dependence bias could not always be reduced to acceptable levels when using only three indicators.

Conclusions:

• The developed modeling strategy is effective in mitigating local dependence bias in latent class analysis, particularly for sensitive survey data.
• Indicator quality is a critical factor influencing the success of bias reduction techniques in LCA.
• While the strategy improves accuracy, careful consideration of indicator selection and model diagnostics is essential for reliable survey estimates.